Multisensor Damage Localization in Buildings via Self-Supervised Vibration Representation Learning

Abstract

Structural Health Monitoring (SHM) has become a critical component in the lifecycle management of civil infrastructure, particularly for high-rise buildings subjected to aging and environmental stressors. Traditional damage identification methods largely rely on supervised learning paradigms that require extensive labeled datasets of damaged states. However, in real-world scenarios, data representing structural failure is sparse, expensive to acquire, and often unavailable until a catastrophic event occurs. This creates a significant bottleneck in deploying data-driven diagnostic systems. To address this label scarcity, this paper proposes a novel framework for Multisensor Damage Localization using Self-Supervised Learning (SSL). We introduce a spatiotemporal graph contrastive learning architecture that exploits the inherent topology of sensor networks and the temporal consistency of vibration responses. By treating the sensor network as a graph and the time-series vibration data as node attributes, our model learns robust, damage-sensitive representations from normal operational data alone. We employ a specialized augmentation strategy tailored for vibration signals, including phase shifting and stochastic sensor masking, to train the network to distinguish between environmental variability and structural anomalies. Experimental validation is conducted on a high-fidelity finite element model of a ten-story building under various excitation profiles. Results demonstrate that the proposed method significantly outperforms traditional autoencoder-based approaches and achieves localization accuracy comparable to fully supervised baselines, utilizing only 5% of the labeled data for fine-tuning.